Abstract: In human body modeling system via stereo vision,removing background pixels could reduce the computational costs of stereo matching and improve human model reconstruction efficiency.For this purpose,an end-to-end deep learning network is proposed to automatically estimate Alpha mask of human body foreground in captured images,where a large number of ground-truth Alpha masks of human body foreground are given.The proposed network includes two stages:human body foreground segmentation stage and Alpha matting of human body foreground stage.In the first stage,a tailored Mask R-CNN,where its object detection and mask regression are reserved,is fined-tuned with the prepared training data to generate a binary mask of the human body foreground.In the second stage,an Encoder-Decoder network is used to address the task of Alpha matting.First,a non-trainable convolutional layer with a kernel size of 5 is applied to the binary mask to generate a Trimap of human body foreground as an input of this stage.After iteratively training with the Trimap data and the human body foreground data,the network can estimate the human foreground alpha mask.In experimental part,taking an acquisition image of human body as an input,the proposed network could output the foreground Alpha matting result fully automatically.Experimental results and detailed comparison with other commonly used matting algorithms are provided to demonstrate accuracy and robustness of the proposed method from both qualitative and quantitative aspects.Meanwhile,the proposed algorithm is also applied to some public data sets.The matting result demonstrates that the proposed method could deal with general images containing human body foreground.